Digital Twin Based Fault-Tolerance Framework for RRAM Based Neural Computing Systems

An RRAM-based computing system (RCS) is a cutting-edge hardware platform that shows significant promise for implementing vector-matrix multiplication in analog form for neuromorphic computing, which can significantly reduce energy consumption. However, one major issue is that current fabrication processes lead to high rates of hard faults and limited endurance of RRAMs, which can cause a drastic reduction in the accuracy and lifespan of RCS. Moreover, training large neural networks on modern computing systems can be time-consuming, further impeding progress in the field. To overcome these obstacles, we introduce a digital-twin-based framework for RCS that is both cost-effective and fault-tolerant. This approach utilizes the natural fault-tolerance of neural networks to with-stand faults during subsequent training phases without requiring additional hardware. To facilitate faster processing, we have devised a solution that integrates fault-tolerance algorithms such as significant classification and remapping, transfer learning, and threshold training into the virtual twin of the digital-twin. To assess the efficacy of our proposed solution, we conducted tests on the CIFAR-10 dataset using VGG16, Alexnet, and ResNet20 networks. The results demonstrate that our proposed framework significantly reduces training time by 90% when compared to current state-of-the-art methods. Moreover, our framework is scalable, secure, and offers 55% lower Total Cost of Ownership (TCO), making it an attractive solution for the future development of RCS.